Most winery owners I talk with about a failed AI rollout describe the same scene. They paid for licenses six months ago. They sent an internal email announcing the rollout. They held a training session, maybe two. Then they checked the admin dashboard a quarter later and found that usage was hovering somewhere between flat and embarrassing. The story that gets told in the next budget meeting is that the team did not take to it.
The story is almost always wrong. The team took to AI months before the winery did. They just took to it on a different tab.
The MIT NANDA group, in its August 2025 GenAI Divide report, put one number against another that should reshape how a winery thinks about its AI investment. Ninety percent of workers say they use personal AI tools every day for personal tasks. Forty percent of firms have an official AI subscription in place. The gap between those two numbers, in MIT's own framing, is the largest single predictor of stalled enterprise AI adoption. They have a name for it. They call it shadow AI.
What shadow AI looks like in a tasting room
Shadow AI is mundane. The wine club concierge writes thank-you notes to a tricky member on her phone during her lunch break, pastes the rough draft into ChatGPT, gets back something cleaner, copies it into Outlook on her work laptop. The tasting room manager asks Claude to sort through four reservation conflicts on a Saturday morning, on her personal phone, while she sips coffee in the parking lot before her shift. The DTC coordinator dictates an apology to ChatGPT in the car after a wholesaler call went sideways.
None of those uses appear on the company dashboard. The winery is paying for an enterprise seat that sits unused. The employee is using a free or twenty-dollar consumer tier that she pays for herself, because that is the tab her phone already had open. The work product lands in the inbox. The owner sees a polished email and assumes the staffer just had a good morning.
The reason the official rollout looks like a flop is that the leader is checking the wrong tab. Real usage on the team has been there for months. It just sits one screen over from where the admin metrics live.
The math behind the 90 versus 40
The ninety-percent figure comes from worker surveys MIT NANDA aggregated across enterprises ranging from twenty-person firms to global ones. Forty percent refers to organizations with an official LLM subscription in place. The middle of that gap, roughly half of all workers, is using AI on the job with a tool the employer did not provide. They are doing it because the consumer tools are good, cheap or free, and already on the phone she uses for personal email between shifts.
Two consequences follow for a small-to-mid-size wine business. The first is that the team's distrust of AI is far smaller than the leader thinks. Most leaders walk into the rollout assuming a long stretch of skepticism and resistance to manage. The actual situation, in most wineries I have looked at, is that the staff has been quietly using AI for personal email and travel planning for a year. They already trust the tool to write a thank-you note. They are waiting for the winery to catch up on which of their tasks at work it is acceptable to point the tool at.
The second consequence is that "we cannot get them to use it" almost never describes a usage problem. It describes a transfer problem. The use already exists. The work has not moved across yet.
Why the official rollout looks dead when it is not
The story leaders tell themselves about a failing rollout follows a predictable shape. We bought licenses. We held training. We posted prompts. The team is not engaging. Maybe they need more training. Maybe AI is not a fit for wine.
The shadow AI data reshapes that story. The team is engaging. They are engaging with consumer tools, on personal accounts, for tasks that happen to include some of the work the winery pays them for. The license sits idle because the staff has no reason to switch tabs. Their personal Claude has their writing samples already pasted in. Their personal ChatGPT remembers that they prefer short paragraphs and no exclamation marks. The winery's enterprise tool starts from a blank screen every time.
A blank screen is the silent killer of an AI rollout. The earlier post in this series walked through MIT's finding that the pilots that survived past month two were the ones where the model had memory of yesterday and the team's work. Consumer accounts already have that memory, accumulated quietly over months of personal use. Enterprise rollouts start over every Monday. The staffer chooses the tool with the memory, every time.
What the leader can do without buying anything else
The move that closes the shadow AI gap costs nothing. There is no new subscription to buy and no memo about acceptable use to circulate. The move is a fifteen-minute conversation with one staffer, and it has three parts.
First, ask what she is already using AI for in her personal life. Some staffers will say nothing. Most will say something specific: a vacation itinerary for her kid's spring break, a cover letter for her sister, a recipe modification, a complicated email to her landlord. Listen for which tasks she already trusts the tool with. That tells you which work tasks she will trust it with first.
Second, ask whether any of those uses overlap with what she does at the winery. The wine club concierge who used AI to write a careful note to her sister already has the muscle for writing a careful note to a club member. The tasting room manager who used it to draft a tricky email to her landlord already has the muscle for a tricky email to a wholesaler. The transfer is closer than it looks.
Third, ask if she would like to point the same tool at one task at work for a couple of weeks, see how it goes, and tell you what she finds. Pick a recurring task, not a one-off. The post-visit follow-up. The thank-you note to a club member who upgraded. The shipment problem reply. Something she does every week and would happily not do.
That is the rollout. No internal email goes out. No training session goes on the calendar. There is a staffer who already trusts the tool, doing a task she already does, with a slightly cleaner workflow than she had on Monday.
The rest of the play, the one I have published as a 60-day pattern, is built around what happens after that fifteen minutes. The short version: get her using the tool on that task three times a week for eight weeks, with a folder full of her real writing samples so the tool stops starting from a blank screen every Monday. Once she has a habit, add a second staffer who does a different kind of work but trusts the first one's judgment.
The fictional winery example
To make this concrete, here is what it looks like for Not Really Wines, our fictional demo winery, which we run as a working showcase of the kind of vault and workflow setup we build for real wineries. (notreallywines.vercel.app)
Priya, NRW's DTC and Society director, started using ChatGPT on her phone in late 2024 to draft holiday email copy for her parents' small business. By spring 2025 she was using it on her personal account for two or three work tasks a week: a quick subject-line variation, a tightened-up paragraph, an outline for a new member welcome. None of that showed up in NRW's books because NRW had not yet bought a license. The work showed up. The tool's involvement did not.
When NRW set up its first official rollout, Maren and Priya skipped the license-and-announce step. Instead, they sat together for ninety minutes and built a folder containing Priya's role description, eight or ten samples of her real club emails, and three example inputs and outputs for the welcome-note task she had been quietly using ChatGPT for already. The folder put her shadow usage on the table. The license came next, sized to where the work already was.
The number of new uses on Priya's part, in the first thirty days, was small. She was already using the tool. What changed was that the work flowed through a workspace the winery could see, on a folder that other staffers could borrow, with a writing voice that started to compound across the team. Tessa, who runs hospitality, asked Priya a question about how she got the tool to sound like her. Two weeks later Tessa had her own folder. That is how the second user enters the play.
What the dashboard cannot see, and what to track instead
A winery owner who wants to know whether AI is working at the business should stop looking at the seat-utilization dashboard. The seat dashboard is set up to measure the wrong thing. It measures whether the employee is logging into the tool the winery pays for. It does not measure whether the employee is using AI at all. For a winery in the first six months of a rollout, those two things are almost never the same.
Track three things instead. None of them need a vendor dashboard.
Track which named staffer is using the tool on a named recurring task, three times a week. One name, one task, three sessions. If the answer at day thirty is zero, the rollout has not started yet, regardless of what the seat metrics say. If the answer is one staffer, one task, three sessions, the play is alive.
Track one time-saved story per active staffer at day sixty. The staffer should be able to say something specific. The post-visit follow-up used to take twenty-five minutes and now takes five. The Saturday morning reservation-conflict scan used to take half an hour and now takes ten. The time-saved story is the unit of evidence that lasts after the dashboard is closed.
Track who is asking the active staffer for help. The Hartz AI research on champion programs found that organic peer pull is the strongest predictor of sustained adoption. If Tessa is asking Priya how she gets the tool to sound like her, the rollout is healthy. If nobody is asking anybody anything, the rollout is still inside the champion bubble and needs a second user installed by week six.
Those three signals are visible without a license dashboard and without a survey. They are visible by walking around.
A note on the webinar
I am running a webinar for wine industry leaders trying to move AI work off personal phones and onto the winery's books. The first half is a live demo of the fifteen-minute conversation and the ninety-minute folder build for a frontline winery role. The second half walks through the research-backed 60-day play that turns one staffer's shadow usage into a real, visible team habit. If your winery has a license already and the dashboard is flat, this is the hour. Check the webinar schedule for the next date and to register.
Part of a 20-post series on employee AI adoption for wineries — see the full series under AI Adoption.